Lecture 13: k-means and

mean-shift clustering
Juan Carlos Niebles
Stanford AI Lab

Professor Fei-Fei Li
Stanford Vision Lab

Fei-Fei Li Lecture 13 - 1 8-Nov-2016

 Recap: Image Segmentation
• Goal: identify groups of pixels that go together

Fei-Fei Li Lecture 13 - 2 8-Nov-2016

 Recap: Gestalt Theory
• Gestalt: whole or group (German: "shape, form")
– Whole is other than sum of its parts
– Relationships among parts can yield new properties/features

• Psychologists identified series of factors that predispose set of

elements to be grouped (by human visual system)
“I stand at the window and see a house, trees, sky.
Theoretically I might say there were 327 brightnesses
and nuances of colour. Do I have "327"? No. I have sky, house,
and trees.”
Max Wertheimer
(1880-1943)

Untersuchungen zur Lehre von der Gestalt,

Psychologische Forschung, Vol. 4, pp. 301-350, 1923

Fei-Fei Li Lecture 13 - 3 8-Nov-2016

 Recap: Gestalt Factors

● These factors make intuitive sense, but are very difficult to translate into algorithms.

Fei-Fei Li Lecture 13 - 4 8-Nov-2016

 Recap: Multistability

https://en.wikipedia.org/wiki/Spinning_Dancer

Fei-Fei Li Lecture 13 - 5 8-Nov-2016

 Recap: Agglomerative clustering

Simple algorithm

● Initialization:
○ Every point is its own cluster
● Repeat:
○ Find “most similar” pair of clusters
○ Merge into a parent cluster
● Until:
○ The desired number of clusters has been reached
○ There is only one cluster

Fei-Fei Li Lecture 13 - 6 8-Nov-2016

 What will we learn today?
• K-means clustering
• Mean-shift clustering

Reading material:
Forsyth & Ponce: Chapter 9.3
Comaniciu and Meer, Mean Shift: A Robust Approach toward Feature Space Analysis, PAMI 2002.
gifs: https://www.projectrhea.org

Fei-Fei Li Lecture 13 - 7 8-Nov-2016

 Image Segmentation: Toy Examplewhite
pixels
3 black pixels
gray

1 2 pixels

input image
• These intensities define the three groups.
intensity
• We could label every pixel in the image according to which
of these primary intensities it is.
– i.e., segment the image based on the intensity feature.
• What if the image isn’t quite so simple?

Slide credit: Kristen Grauman

Fei-Fei Li Lecture 13 - 8 8-Nov-2016

 Pixel count
Input image
Intensity

Pixel count

Input image
Intensity
Slide credit: Kristen Grauman
Fei-Fei Li Lecture 13 - 9 8-Nov-2016
 Pixel count
Input image
Intensity

• Now how to determine the three main

intensities that define our groups?
• We need to cluster.

Slide credit: Kristen Grauman

Fei-Fei Li Lecture 13 - 10 8-Nov-2016

 0 190 255
Intensity

3
1 2

• Goal: choose three “centers” as the representative intensities,

and label every pixel according to which of these centers it is
nearest to.

Slide credit: Kristen Grauman

• Best cluster centers are those that minimize Sum of Square
Distance (SSD) between all points and their nearest cluster
center ci:

Fei-Fei Li Lecture 13 - 11 8-Nov-2016

 Objective function
● Goal: minimize the distortion in data given
clusters
- Preserve information
Cluster center Data

Slide: Derek Hoiem

Fei-Fei Li Lecture 13 - 12 8-Nov-2016
 Clustering
• With this objective, it is a “chicken and egg”
problem:
– If we knew the cluster centers, we could allocate points to
groups by assigning each to its closest center.

Slide credit: Kristen Grauman

– If we knew the group memberships, we could get the
centers by computing the mean per group.

Fei-Fei Li Lecture 13 - 13 8-Nov-2016

 K-means Clustering
● Initialization:
○ choose k cluster centers
● Repeat:
○ assignment step:
■ For every point find its closest center
○ update step:
■ Update every center as the mean of its points
● Until:
○ The maximum number of iterations is reached, or
○ No changes during the assignment step, or
○ The average distortion per point drops very little

[Lloyd, 1957]

Fei-Fei Li Lecture 13 - 14 8-Nov-2016

 K-means Clustering te c h n ica l
note

slide credit: P. Rai

Fei-Fei Li Lecture 13 - 15 8-Nov-2016

 K-means Clustering te c h n ica l
note

[1]

[1] L. Bottou and Y. Bengio. Convergence properties of the kmeans algorithm. NIPS, 1995. slide credit: P. Rai

Fei-Fei Li Lecture 13 - 16 8-Nov-2016

 K-means: Initialization
● k-means is extremely sensitive to initialization
● Bad initialization can lead to:
○ poor convergence speed
○ bad overall clustering
● How to initialize?
○ randomly from data
○ try to find K “spread-out” points (k-means++)
● Safeguarding measure:
○ try multiple initializations and choose the best

Fei-Fei Li Lecture 13 - 17 8-Nov-2016

 K-means: Initialization
● k-means is extremely sensitive to initialization
● Bad initialization can lead to:
○ poor convergence speed
○ bad overall clustering

Fei-Fei Li Lecture 13 - 18 8-Nov-2016

 K-means++
• Can we prevent arbitrarily bad local minima?

1. Randomly choose first center.

2. Pick new center with prob. proportional to (x - ci)2
– (Contribution of x to total error)
3. Repeat until K centers.

• Expected error O(logK) (optimal)

K-means++ animation Arthur & Vassilvitskii 2007

Fei-Fei Li Lecture 13 - 19 8-Nov-2016

 K-means: choosing K

slide credit: P. Rai

Fei-Fei Li Lecture 13 - 20 8-Nov-2016

 K-means: choosing K

• Validation set
– Try different numbers of clusters and look at
performance
• When building dictionaries (discussed later), more
clusters typically work better

Slide: Derek Hoiem

Fei-Fei Li Lecture 13 - 21 8-Nov-2016

 Distance Measure & Termination
● Choice of “distance” measure:
■ Euclidean (most commonly used)
■ Cosine
■ non-linear! (Kernel k-means)

Picture courtesy: Christof Monz (Queen Mary, Univ. of London)

● Termination:
○ The maximum number of iterations is reached
○ No changes during the assignment step (convergence)
○ The average distortion per point drops very little

Fei-Fei Li Lecture 13 - 22 8-Nov-2016

 K-means: Example

K-Means Clustering Example

Fei-Fei Li Lecture 13 - 23 8-Nov-2016

 How to evaluate clusters?
• Generative
– How well are points reconstructed from the clusters?
→ “Distortion”

• Discriminative
– How well do the clusters correspond to labels?
• Purity
– Note: unsupervised clustering does not aim to be
discriminative

Slide: Derek Hoiem

Fei-Fei Li Lecture 13 - 24 8-Nov-2016
 Segmentation as Clustering
● Let’s just use the pixel intensities!

k=2

k=3

Slide credit: Kristen Grauman

Fei-Fei Li Lecture 13 - 25 8-Nov-2016

 Feature Space
• Depending on what we choose as the feature space,
we can group pixels in different ways.

• Grouping pixels based on

intensity similarity

• Feature space: intensity value (1D)

Slide credit: Kristen Grauman

Fei-Fei Li Lecture 13 - 26 8-Nov-2016

 Feature Space
• Depending on what we choose as the feature space, we can
group pixels in different ways.
R=255
• Grouping pixels based G=200
on color similarity B=250

R=245
B G=220
G B=248

R=15 R=3
R G=189 G=12
B=2 B=2

● Feature space: color value (3D)

Slide credit: Kristen Grauman

Fei-Fei Li Lecture 13 - 27 8-Nov-2016

 Feature Space
• Depending on what we choose as the feature space, we can
group pixels in different ways.
• Grouping pixels based
on texture similarity
F1
F2
Filter bank of
24 filters
…

F24

● Feature space: filter bank responses (e.g., 24D)

Slide credit: Kristen Grauman

Fei-Fei Li Lecture 13 - 28 8-Nov-2016

 K-Means Clustering Results
• K-means clustering based on intensity or color is
essentially vector quantization of the image attributes
– Clusters don’t have to be spatially coherent
Image Intensity-based clusters Color-based clusters

Image source: Forsyth & Ponce

Fei-Fei Li Lecture 13 - 29 8-Nov-2016
 Smoothing Out Cluster Assignments
• Assigning a cluster label per pixel may yield outliers:

Original Labeled by cluster center’s intensity

?
• How can we ensure they
are spatially smooth? 3
1 2

Slide credit: Kristen Grauman

Fei-Fei Li Lecture 13 - 30 8-Nov-2016

 Segmentation as Clustering
• Depending on what we choose as the feature space,
we can group pixels in different ways.
• Grouping pixels based on
intensity+position similarity
Intensity

X
⇒ Way to encode both similarity and proximity.
Slide credit: Kristen Grauman

Fei-Fei Li Lecture 13 - 31 8-Nov-2016

 K-means clustering for superpixels
SLIC Superpixels:
● Feature space → intensity + position
○ L.a.b. color space
○ limited region (window 2*S)

● Distance metric:

● Initialization:
○ Spatial grid (grid step = S)
● Iterate over centers and not points
Achanta et al., SLIC Superpixels Compared to State-of-the-art Superpixel Methods, PAMI 2012.

Fei-Fei Li Lecture 13 - 32 8-Nov-2016

 K-means clustering for superpixels

Achanta et al., SLIC Superpixels Compared to State-of-the-art Superpixel Methods, PAMI 2012.
Fei-Fei Li Lecture 13 - 33 8-Nov-2016
 K-means clustering for superpixels

Achanta et al., SLIC Superpixels Compared to State-of-the-art Superpixel Methods, PAMI 2012.
Fei-Fei Li Lecture 13 - 34 8-Nov-2016
 K-means clustering for superpixels

Achanta et al., SLIC Superpixels Compared to State-of-the-art Superpixel Methods, PAMI 2012.
Fei-Fei Li Lecture 13 - 35 8-Nov-2016
 K-means Clustering: Limitations

[1]

[1] Dhillon et al. Kernel k-means, Spectral Clustering and Normalized Cuts. KDD, 2004. slide credit: P. Rai

Fei-Fei Li Lecture 13 - 36 8-Nov-2016

 K-Means pros and cons
• Pros
• Finds cluster centers that minimize
conditional variance (good
representation of data)
• Simple and fast, Easy to implement
• Cons
• Need to choose K
• Sensitive to outliers
• Prone to local minima
• All clusters have the same
parameters (e.g., distance measure
is non-adaptive)
• *Can be slow: each iteration is
O(KNd) for N d-dimensional points
• Usage
• Unsupervised clustering
• Rarely used for pixel segmentation

Fei-Fei Li Lecture 13 - 37 8-Nov-2016

 Scaling-up K-means clustering
● Assignment step is the bottleneck

● Approximate assignments
○ [AK-means, CVPR 2007], [AGM, ECCV 2012]
● Mini-batch version
○ [mbK-means, WWW 2010]
● Search from every center
○ [Ranked retrieval, WSDM 2014]
● Binarize data and centroids
○ [BK-means, CVPR 2015]
● Quantize data
○ [DRVQ, ICCV 2013], [IQ-means, ICCV 2015]

Fei-Fei Li Lecture 13 - 38 8-Nov-2016

 What will we learn today?
• K-means clustering
• Mean-shift clustering

Fei-Fei Li Lecture 13 - 39 8-Nov-2016

 Mean-Shift Segmentation
• An advanced and versatile technique for
clustering-based segmentation

Slide credit: Svetlana Lazebnik

D. Comaniciu and P. Meer, Mean Shift: A Robust Approach toward Feature Space Analysis, PAMI 2002.

Fei-Fei Li Lecture 13 - 40 8-Nov-2016

 Mean-Shift Algorithm

Slide credit: Steve Seitz

• Iterative Mode Search
1. Initialize random seed, and window W
2. Calculate center of gravity (the “mean”) of W =
3. Shift the search window to the mean
4. Repeat Step 2 until convergence

[Fukunaga & Hostetler, 1975]

Fei-Fei Li Lecture 13 - 41 8-Nov-2016
 Mean-Shift Region of
interest

Center of
mass

Mean Shift
vector
Slide by Y. Ukrainitz & B. Sarel

Fei-Fei Li Lecture 13 - 42 8-Nov-2016

 Mean-Shift Region of
interest

Center of
mass

Mean Shift
vector
Slide by Y. Ukrainitz & B. Sarel

Fei-Fei Li Lecture 13 - 43 8-Nov-2016

 Mean-Shift Region of
interest

Center of
mass

Mean Shift
vector
Slide by Y. Ukrainitz & B. Sarel

Fei-Fei Li Lecture 13 - 44 8-Nov-2016

 Mean-Shift Region of
interest

Center of
mass

Mean Shift
vector
Slide by Y. Ukrainitz & B. Sarel

Fei-Fei Li Lecture 13 - 45 8-Nov-2016

 Mean-Shift Region of
interest

Center of
mass

Mean Shift
vector
Slide by Y. Ukrainitz & B. Sarel

Fei-Fei Li Lecture 13 - 46 8-Nov-2016

 Mean-Shift Region of
interest

Center of
mass

Mean Shift
vector
Slide by Y. Ukrainitz & B. Sarel

Fei-Fei Li Lecture 13 - 47 8-Nov-2016

 Mean-Shift Region of
interest

Center of
mass

Slide by Y. Ukrainitz & B. Sarel

Fei-Fei Li Lecture 13 - 48 8-Nov-2016

 Real Modality Analysis

Slide by Y. Ukrainitz & B. Sarel

Tessellate the space with windows Run the procedure in parallel
Fei-Fei Li Lecture 13 - 49 8-Nov-2016
 Real Modality Analysis

Slide by Y. Ukrainitz & B. Sarel

The blue data points were traversed by the windows towards the mode.
Fei-Fei Li Lecture 13 - 50 8-Nov-2016
 Mean-Shift Clustering
• Cluster: all data points in the attraction basin of a
mode
• Attraction basin: the region for which all trajectories
lead to the same mode

Slide by Y. Ukrainitz & B. Sarel

Fei-Fei Li Lecture 13 - 51 8-Nov-2016

 Mean-Shift Clustering/Segmentation
• Find features (color, gradients, texture, etc)
• Initialize windows at individual pixel locations
• Perform mean shift for each window until convergence
• Merge windows that end up near the same “peak” or mode

Slide credit: Svetlana Lazebnik

Fei-Fei Li Lecture 13 - 52 8-Nov-2016
 Mean-Shift Segmentation Results

Slide credit: Svetlana Lazebnik

Fei-Fei Li Lecture 13 - 53 8-Nov-2016
 More Results

Slide credit: Svetlana Lazebnik

Fei-Fei Li Lecture 13 - 54 8-Nov-2016
 More Results

Fei-Fei Li Lecture 13 - 55 8-Nov-2016

 Problem: Computational Complexity

Slide credit: Bastian Leibe

• Need to shift many windows…
• Many computations will be redundant.

Fei-Fei Li Lecture 13 - 56 8-Nov-2016

 Speedups: Basin of Attraction

Slide credit: Bastian Leibe

1. Assign all points within radius r of end point to the mode.

Fei-Fei Li Lecture 13 - 57 8-Nov-2016

 Speedups

Slide credit: Bastian Leibe

● Assign all points within radius r/c of the search path to the mode
Fei-Fei Li Lecture 13 - 58 8-Nov-2016
 n i cal
tech e
not

Mean-shift
Algorithm
Comaniciu & Meer, 2002

Fei-Fei Li Lecture 13 - 59 8-Nov-2016

 n i cal
tech e
not

Mean-shift
Algorithm
Comaniciu & Meer, 2002

Fei-Fei Li Lecture 13 - 60 8-Nov-2016

 Summary Mean-Shift
• Pros
– General, application-independent tool
– Model-free, does not assume any prior shape (spherical,
elliptical, etc.) on data clusters
– Just a single parameter (window size h)
• h has a physical meaning (unlike k-means)
– Finds variable number of modes
– Robust to outliers
• Cons

Slide credit: Svetlana Lazebnik

– Output depends on window size
– Window size (bandwidth) selection is not trivial
– Computationally (relatively) expensive (~2s/image)
– Does not scale well with dimension of feature space

Fei-Fei Li Lecture 13 - 61 8-Nov-2016

 Medoid-Shift & Quick-Shift

• Quick-Shift:
- does not need the gradient or quadratic lower bound
- only one step has to be computed for each point: simply moves each
point to the nearest neighbor for which there is an increment of the
density
- there is no need for a stopping/merging heuristic
- the data space X may be non-Euclidean

[Vedaldi and Soatto, 2008]

Fei-Fei Li Lecture 13 - 62 8-Nov-2016

 What have we learned today
• K-means clustering
• Mean-shift clustering

IPython Notebook for SLIC and Quickshift

Fei-Fei Li Lecture 13 - 63 8-Nov-2016